Consider we have trained a classifier C
, for example, a Multilayer Perceptron (MLP), or a Support Vector Machines (SVM), to classify instances from a good dataset E
provided by "Experts". There is another poor dataset A
of the same task generated by "Apprentices" or "Adversarial Attackers", on which the classifier C
performs poorly. The goal of hypothetical refinement is to raise a what if question: "What if these data look like another way?" and to modify the instance to pass classifier C
with high confidence by associative inference.